Remove temporal solution for selecting type of input dataset

[gcerar]

The CTW2019 and CTW2020's dvc.yaml config currently uses a list of models and inputs. This was a temporary workaround when the solution was a work in progress. The final goal is to move all models and input types/variants specified in params.yaml file. I think this still needs to be talked through. A possible solution is something like this:

# dvc.yaml
  benchmark:
    desc: Determine best hyper-parameters for algorithm on several CVs
    matrix:
      split: ${splits}
      model:  # <--- We remove this
        - name: XGBRegressor
          input: ${path.data}/prepared/ctw2020.flatten.pkl
        - name: RandomForestRegressor
          input: ${path.data}/prepared/ctw2020.flatten.pkl
        - name: Arnold2018DeepModel
          input: ${path.data}/prepared/ctw2020.flatten.pkl
        - name: Arnold2019SoundingModel
          input: ${path.data}/prepared/ctw2020.pkl
        - name: Pirnat2022Pirnat1G
          input: ${path.data}/prepared/ctw2020.pkl
        - name: Cerar2021Localization
          input: ${path.data}/prepared/ctw2020.pkl
    ...

# params.yaml
models:
  Arnold2018DeepModel:
    input_type: bchw # <--- and introduce this (possible values something like: flatten, bchw, chwc)
    ... 

It's crucial to update the benchmark.py script to match the changes in the dvc.yaml and params.yaml configs.